Researchers at Pillar Security demonstrated sandbox escapes across four widely used AI coding agents: Cursor, OpenAI Codex CLI, Google Gemini CLI, and Antigravity. In nearly every case the agent never broke the sandbox directly; it only had to write a file that a trusted component outside the sandbox would later run, load, or scan. Failure modes included hook abuse, editing a virtual environment interpreter the editor then ran itself, planting Git metadata outside a .git folder to fire execution through fsmonitor, and a command allowlist that trusted a tool by name while the real invocation was not read only. Prompt injection in workspace content was the trigger.
Researchers described Agent Data Injection, a new twist on prompt-injection attacks against AI agents. Rather than smuggling in fake instructions, it exploits the weak separation between trusted and untrusted data so that attacker-supplied content is mistaken for the agent's own trusted data, using deliberately ambiguous delimiters the model misreads. In tests against web and coding agents, this let an attacker steer an agent's clicks or actions, succeeding up to half the time even against defenses that block ordinary instruction injection. Some approaches helped: tagging page elements with random, unguessable identifiers roughly halved success, while strict tracking of where data came from stopped it but sharply reduced how many tasks agents completed.
Researchers demonstrated MemGhost, an attack that uses one email to plant a false, persistent memory in an AI personal assistant with inbox access. Because these assistants keep notes about the user and reload them every session, a crafted message can trick the agent into saving a fabricated fact while keeping its reply innocuous, so the tampering goes unnoticed and steers later answers. The team trained an attacker model to write such emails automatically, reporting high success against open-source and commercial agents and showing it transfers across memory backends and survives several defenses. Unlike earlier one-shot injection that leaked data only in the moment, MemGhost's memory persists long after the email is gone.
Researchers at the ASSET Research Group demonstrated Ghostcommit, an attack that hides malicious instructions inside a PNG image so AI code reviewers miss them entirely. A harmless-looking conventions file points the coding agent to the image, whose rendered text tells the agent to read the repository's .env file and encode its secrets as a list of numbers written into the code. Because tools like CodeRabbit and Bugbot skip image files by default, the pull request passes review clean; the trap springs later when a developer asks the agent for an unrelated task. The outcome depended on the tool wrapping the model: some agents leaked secrets, while Claude Code refused.
Researchers at Noma Labs showed that GitHub's new Agentic Workflows, which let an AI agent read issues and act on repositories automatically, can be tricked into leaking private code through nothing more than a public issue. The technique, GitLost, is indirect prompt injection: an attacker opens an ordinary-looking issue in an organization's public repository, buries plain-English instructions in it, and the agent, which often holds a token with read access across the org's repositories, follows them, fetches files from a private repo, and posts the contents in a public comment. No credentials, coding, or write access are needed. GitHub was notified, but researchers frame it as an architectural weakness.
Zscaler found attackers using search-engine poisoning and hidden instructions on malicious websites to manipulate AI agents into making cryptocurrency payments. In one case, a hidden element on the page tells an AI agent that it must "resolve an error" by completing a payment, alongside code that starts a crypto transfer to a hardcoded wallet; the same page also shows human visitors ordinary payment options. Another campaign typosquats a decentralized-finance portfolio tracker and uses hidden prompts to convince agents the fake site is the real one. The attacker is seeding the scheme through several GitHub repositories, showing how autonomous agents that browse and act can be steered by content they read.
Researchers at Cato AI Labs detailed two flaws, dubbed DuneSlide, in the AI code editor Cursor that let a prompt-injection attack break out of the sandbox Cursor uses to contain the commands its agent runs. The attacker never types anything: they plant instructions in content the agent reads on the user's behalf, such as a connected MCP service or a web page. One flaw abuses a working-directory setting to get an attacker path added to the allowed-write list, letting injected commands overwrite the sandbox helper itself and then run with no sandbox. Both are rated 9.8 and are fixed in Cursor 3.0; every earlier version is affected, so users should update.
Microsoft is warning that attackers can hijack AI agents through poisoned tool descriptions, the plain-text notes that tell an agent what a tool does. Because agents connect to systems through the Model Context Protocol and read these descriptions to decide how to act, an attacker who updates a trusted third-party tool can bury a hidden instruction in its description, telling the agent to quietly collect and exfiltrate data on its next task. Many setups pick up description changes without re-approval, so the poisoned version goes live silently. Each step the agent takes looks legitimate and runs with the user's own permissions, so no alarm fires.
Researchers at LayerX detailed BioShocking, an attack that manipulates AI browser agents into ignoring their safety rules by convincing them they are inside a fictional game. Using a web page with a puzzle that rewards deliberately wrong answers, the attack gets the agent to accept a false reality, after which it treats a request to open a page and copy its contents as just another step. In the demonstration, that page redirected to the victim's work GitHub repository and the agent handed over SSH credentials, treating the theft as finishing the game. None of the six AI browser agents tested flagged it as a rule violation.
SentinelOne detailed Gaslight, a Rust-based macOS backdoor and information stealer tied with high confidence to North Korea, whose standout trick targets the analyst rather than the sandbox. The sample embeds a block of 38 fabricated "system" messages, formatted to mimic the prompt scaffolding of an AI triage assistant, that try to make an LLM-assisted analysis tool doubt its session and abort, truncate, or refuse the analysis. Beyond that, Gaslight steals browser data, Keychain secrets, and command history, using a Telegram bot for command and control and self-redacting its bot token from its own output. It is an early example of malware built to weaponize the AI tools now common in reverse engineering.
TrustStrike Labs